-
A set of machine learning experiments with the semeion and MNIST handwritten digit dataset using tensorflow
-
The objective of this experiment was to test multiple classification methods using the semeion handwriting dataset and measure performance of different classifiers implementation in tensorflow.
-
The semeion dataset is composed of 1593 handwritten digits from 80 persons that were scanned and stretched to a 16x16 size image.
-
http://archive.ics.uci.edu/ml/datasets/semeion+handwritten+digit
-
MNIST dataset is a subset of the NIST dataset that has over 60000 handwritten digits.
-
To change the dataset, change the dataset loading code and sample size in the implementation files.
-
If you want to you can also import your own dataset, this code can be easily adapted to classify other type of images.
width = 16
height = 16
dataset = semeion.read_data_semeion()
-
The code available was tested with Python 3.5 and Tensorflow 1.1
-
Before running the examples in the repository, install the dependencies.
tensorflow matplotlib sklearn pandas numpy
- Clone the repository into your computer
- Dataset files are already included in the repository inside the /source/dataset folder.
- Run one of the implementation files from the source folder, each one implements a diferent classifier. - knn.py, softmax.py, perceptron.py, cnn.py, lstm.py
- The results bellow were obtained, using 1300 random entries from the semeio dataset to train the classifier and 400 random entries to test the trained model.
- The results obtained are expected, for the recurrent network (long short term memory), i haven't applied any confusion to the input, so after some time it detects that probably the next sample its equal to the current one.
- Tests were run on a Core i5 6500 CPU with 24GB of RAM.
| Classifier | Time | Accuracy |
|---|---|---|
| Softmax | 40.8 | 94.32% |
| KNN | 3.85 | 94.52% |
| Perceptron | 11.4 | 97.16% |
| CNN | 89.7 | 96.95% |
| RNN | 35.1 | 97.56% |